H3 Combat Scenes — Fight Scene Workflow Pack - For Patreon Supporters (Patreon)

Public Post about this LoRA and Diffusion Model

https://www.patreon.com/aifuturetech/posts/minimax-h3-fast-168950662

Hey supporters. This pack matches the workflows I re-run in the new fight scenes video, and this post is the full run manual:

what to download, what to type, what to set. You should be able to reproduce everything without scrubbing the video.

Two workflows, both updated to my current setup:

1. MinimaxH3_t2v_i2v-3S-20260907.json — text-to-video + image-to-video (start/end frames)

2. MinimaxH3_R2V-3S-20260907.json — reference-to-video, character sheets, up to 5 reference inputs

They run different bases on purpose. The t2v/i2v workflow uses the standard MiniMax H3 fl2va model with the LoRA stack. The R2V workflow now runs the Singularity fine-tune (v1.3) as the base model, with VDN-H3 for the fast sampling — that's the combination that won the comparison in the video, so it became my default for character fight scenes.

LoRAs

https://huggingface.co/Jojocodex/wushu-action-v7-minimax-h3-fl2va-ref2va-lora/

https://huggingface.co/JOKER141/MiniMax-H3-Combat-Base-V2

https://huggingface.co/JOKER141/MiniMax-H3-General-Motion-Continuity-Repair

Or

Diffusion Model - Fine Tune

https://huggingface.co/WarmBloodAban/Minimax-h3_Singularity

2 DLSS5 Nodes for options:

https://github.com/Blueforcer/ComfyUI-DLSS5-Enhancer#installation

https://github.com/Konohamaru04/ComfyUI-NVIDIA-DLSS-Frame-Interpolation

----------------------------------------

WHAT TO DOWNLOAD (MODEL FILES)

----------------------------------------

My subfolder is named "minimax-h3" under models — keep the same names or re-point the loaders.

Base diffusion models:

minimax-h3/Minimax-h3_Singularity_ref2va_v1.3_int8.safetensors (R2V workflow base)

minimax-h3/minimax_h3_fl2va_int8_convrot.safetensors (t2v/i2v workflow base)

Text encoder:

minimax-h3/qwen3vl_32b_minimax_h3_int8_convrot.comfy.safetensors

VAEs:

minimax-h3/minimax_h3_video_vae_fp16.safetensors

minimax-h3/minimax_h3_audio_vae_fp32.safetensors

LoRAs (links in the YouTube description):

minimax-h3/wushu_action_v7_fl2va_aitoolkit_adaln_full-int8convrot_bf16te_2000step.safetensors

minimax-h3/H3_Combat_V2.safetensors

minimax-h3/Motion_Repair.safetensors

minimax-h3/minimax_h3_fl2v_turbo_4step_v1.1_768p_comfyui_bf16.safetensors (Turbo)

minimax-h3/h3-realism-people-t2v-i2v-r2v(r34l1sm).safetensors (realism polish)

VDN-H3 checkpoint (R2V workflow only):

The stage-dmd-step-250 folder from the official VDN Hugging Face repo, the whole

folder with all files, saved into your models folder as "vdn". Same install as

the Video DeltaNet video.

Frame interpolation:

film_net_fp16.safetensors

TRT VAE engines (optional, NVIDIA only):

minimax-h3/minimax_h3_vae_decoder.engine

minimax-h3/minimax_h3_vae_encoder.engine

Don't have them? The workflow ships a compiler node that builds them from the

ONNX files. On AMD or non-TRT setups, bypass the TRT loader and connect the

fp16 video VAE directly — everything else runs the same, decode is just slower.

----------------------------------------

QUICK START

----------------------------------------

1. Load a workflow. t2v/i2v if you're prompting text only; R2V if you have character sheets.

2. Check every loader points at your files. The dropdowns remember my paths, not yours.

3. Type your prompt in the big text box. See the next section — and note you do NOT need to type the trigger words yourself, the workflow prepends them automatically.

4. Set Duration and FPS. R2V ships at 15 seconds, t2v/i2v at 10. FPS is 24. Frame count math is handled internally so the clip length lands on a count H3 likes.

5. Set Resolution (megapixels input). R2V ships at 0.8, t2v/i2v at 0.6. Aspect locks to 16:9. Handy conversions at multiples of 32:

0.6 MP = 1056x608 0.8 MP = 1216x672 0.98 MP = 1344x768 1.2 MP = 1504x832

6. Seed. t2v/i2v randomizes by default. R2V ships fixed — that's how I keep comparison runs reproducible. Set it back to randomize for normal exploring.

7. Queue.

----------------------------------------

WHAT TO INPUT (PROMPT FORMAT)

----------------------------------------

The trigger words are handled for you now — both workflows concatenate this prefix onto whatever you type:

r34l1sm , ultra realistic video. wushu_action , Combat Scene.

So you just write the fight. Structure that works best with the wushu LoRA's training format:

Line 1: overall style, e.g. "cinematic hardcore realistic combat, no beauty filter" — the no-filter part stops everyone posing for camera.

Block 2: character setup, one line each: attacker or defender, hair, outfit colors, weapon, powers or none. Keep it factual.

Block 3: engagement rule, one line: how the fight starts and that every hit gets a reaction.

Block 4: shot list, [Shot 1] to [Shot N], one line each: camera move + who does what. ONE action per shot. Name the camera (chest-mounted handheld follow, low tracking, quick push-in on impact).

Block 5: soundscape, one line — this model generates audio, so this feeds the soundtrack. Music line: leave it None unless you want score.

Feeding images:

- t2v: nothing extra, just the prompt.

- i2v: load your start frame (and optional end frame) in the image loaders. Mid-motion frames work better than standing poses — the model continues energy better than it invents it.

- R2V: load character sheet images into the reference inputs — it takes up to 5, resolution set to match, padding handled automatically. Character setup lines in your prompt should match who's in the sheets.

----------------------------------------

KEY SETTINGS (AND WHY)

----------------------------------------

Sampling — two stages per workflow, and they differ:

t2v/i2v: stage 1 = euler, 10 steps, denoise 1.0

stage 2 = euler, 4 steps, denoise 0.35

R2V: stage 1 = res_multistep, 8 steps, denoise 1.0 (VDN-H3 handles the speed here)

stage 2 = euler, 4 steps, denoise 0.45

Why two stages: one pass alone gives soft texture on fast action. The short second pass at low denoise sharpens bodies and weapons without changing the motion — same trick since AnimateDiff. Stage 2 denoise above 0.5 invents new motion and you get flicker. Stay at 0.35-0.45.

LoRA strengths:

t2v/i2v: wushu_action_v7 0.7 / Combat V2 0.5 / Motion Repair 0.5 / Turbo 1.0 / realism 0.7 (stage 2 only)

R2V: wushu_action_v7 0.73 / Combat V2 0.5 / Motion Repair 0.5 / Turbo 1.0 / realism 0.9 (stage 2 only)

The realism LoRA is wired into the refine stage only — it polishes skin and fabric without touching the stage-1 choreography. That's why it can run hotter (0.7-0.9) than the motion LoRAs. Combat V2 and Motion Repair stay at 0.5 so the fight motion stays in charge; 0.6 hits harder but past that the physics fights the choreography.

Attention and memory:

VDN-H3 (R2V only): stage-dmd-step-250, this is what makes 8-step sampling viable on the Singularity base.

Sage Attention: auto, memory-efficient patch on. Leave as shipped.

Chunked feed-forward: chunk 2, size 4096 — the low-VRAM helper. Barely costs speed, keeps memory flat on long clips. Leave it on.

Sigma shift: 12, 3 on both stages, as shipped. Don't touch unless experimenting.

Stage 3 (the polish group, both workflows):

DLSS 5 Enhance Images: on, 1.5x Quality — NVIDIA RTX upscaling on the decoded frames.

FILM frame interpolation: 2x — doubles 24fps output to ~48fps smoothness.

AMD/older cards: bypass this group; you get the clean 2-stage result without polish.

Output: h264 MP4, CRF 19, 24fps, generated audio muxed in. The "vfi" output is the 2x-interpolated copy.

----------------------------------------

TIPS

----------------------------------------

1. Write shots like a fight choreographer, not a poet. "Both swords slash into his chest, he's sent flying along the line of force" gives the physics LoRA something to chew. "An epic battle happens" gives it nothing.

2. One action per shot line. Duck, grab, and throw in one shot averages into mush — split them across shots.

3. R2V at 15s / 0.8 MP is heavier than t2v at 10s / 0.6 MP — roughly double the generation time. The chunked feed keeps it from OOMing on 12GB cards, it just gets slow. On 8GB, drop to 0.5 MP and shorter duration first.

4. The reference images get padded, not cropped — so full-body character sheets with some margin survive better than tightly cropped ones.

5. Generated impact audio is usable but rough on big hits. I keep the ambience (wind, whooshes) and re-create the heavy impacts in the edit.

6. Fixed seed on R2V is a feature: change one thing at a time (a shot line, a strength) and the comparison stays honest.

----------------------------------------

COMMON ISSUES

----------------------------------------

Legs or arms bend wrong mid-clip

→ Motion Repair too low, or a shot line crams two actions together. Raise Motion Repair toward 0.6 and split the action.

Impacts look weak, characters phase through hits

→ Combat V2 missing or below 0.5, and your shot lines need the hit AND the reaction stated. No reaction line, no reaction.

Prompt ignored, characters pose for camera

→ The trigger prefix should be automatic — check the concatenate node is still connected ahead of your prompt. Also keep the "no beauty filter" style line; without it everyone starts posing.

Pixelated dust or smear on fast motion

→ That's what stage 2 and the DLSS 5 pass exist for. If it survives both, your stage 2 denoise drifted below 0.35 — set it back.

OOM on longer generations

→ Confirm the chunk feed node (2 / 4096) is connected, close VRAM eaters, drop to 0.5 MP before touching anything else.

TRT VAE errors on non-NVIDIA

→ The .engine files are NVIDIA-only. Bypass the TRT group, connect the fp16 video VAE directly.

----------------------------------------

That's the full setup — same stack I run everything through now.

Thanks for supporting the channel. You guys make these deep dives possible.

— Benji